World Models are the Future of Gaming.
- Tweets 62
- Following 71
- Followers 7
- Likes 24
Leave a generated place, then come back.
Did the same place wait—or did the model quietly replace it?
DreamX-World retrieves earlier views to preserve old places. In the authors’ return test, it led 5 of 6 measures—while longer interactions can still drift.
41
Thanks to Black Forest Labs for the Free Flux usage!
FLUX 3 Video is #2 in the world.
To celebrate, FLUX 3 Video is free to use in our playground until Sunday 16th, 11:59pm PT (link in the thread).
This is just the beginning of what’s coming. Up next: 4K, video editing, and multiple images & videos as reference input.
Made with AI
63
A perfect dream can fail.
At τ=0.10, a controller averaged 2,086 of 2,100 steps inside its learned Doom world—but only 193 in the actual game.
Raising model temperature helped transfer, until more uncertainty hurt. From Ha & Schmidhuber’s World Models experiment.
24
All these new video models are going to make the next generation of gaming possible.
Video is 24 frames per second. Once we can double that and make it low-latency, your indie games are going to look just like this.
1
29
Video models no longer make obvious mistakes.
There are signs, but they're subtle.
Overly rigid movement. Hallucinatory frames that force double takes. Impossible spawn positions.
Going forward, the aesthetics and taste of a video model will be all that matters.
threw the newest video models at one 10-second one-shot: a sword duel in a bamboo grove
Wan 3.0, LTX 2.5 Pro, Flux 3, Minimax H3
same prompt for all four
40
I had a question about the distilled model?
Introducing LTX-2.5, the world model the world builds on.
One of the biggest upgrades yet to the model already powering film, robotics, and real-time workflows. Higher pixel fidelity, multishot scenes that hold together across cuts, and a pretrained foundation built to be finetuned across domains.
Plus: Diffusion Fidelity Rendering, a new rendering approach built for pixel quality that holds up frame by frame, even on a cinema screen.
This isn't a tool you rent. It's a foundation you build on, open and yours.
52
No control log. Still controllable.
CADDY learns user-selectable actions from tennis video without frame-level action labels—then uses each chosen action to generate the next frame.
Controllable video, not a full game engine.
Paper: arxiv.org/abs/2101.12195
1
44
10 Minecraft training seeds. Every one found at least one diamond within its own 100M-step budget.
DreamerV3’s result is striking—but diamonds still appeared in just 0.4% of episodes. Benchmark discovery, not routine mastery.
27
87 hours of Counter-Strike recordings became a world you can play.
DIAMOND is a diffusion world model that generates the next frame after every action—and shows why visual detail can change what an agent learns.
35
Don’t watch just one window. Watch whether all four agree.
MIRA takes each car’s steering, jump, and boost commands, then generates the next moments of the same Rocket League play across four views.
One shared world, seen from four perspectives.
35
This isn’t Minecraft. It’s an AI generating a playable Minecraft world in real time—one frame at a time, based on your controls.
It’s called OASIS, and it’s a glimpse of what games could look like without a traditional game engine.
36
